Summarizing non-stationarity in spatio-temporal neural data

The brain is exceptional at responding to changes in its environment: a fast-moving object, a new sound, or a demanding task. Investigating how environmental parameters translate to complex, high-dimensional neuronal activity requires powerful techniques that can uncover patterns buried in large spatio-temporal datasets. Standard methods of dimensionality reduction excel at finding global low-dimensional structure in multivariate datasets, but do not provide fine-grained information about how the dynamics of individual neural elements change over time. Such local dynamics can be crucial for understanding the function and organization of heterogeneous structures in the brain. In this talk, I will present a new method for producing compact and interpretable summaries of non-stationary time series. By representing time series as trajectories in a low-dimensional feature space, where each dimension quantifies an informative dynamical property, we can directly observe non-stationary variation in neural activity. Using this method, we are able to reduce the dimensionality of spatio-temporal data without losing a local perspective on dynamics. I will demonstrate our novel approach using the Allen Neuropixels Visual Coding dataset, a highly spatio-temporal set of electrophysiological recordings from the mouse brain aimed at investigating neural responses to visual stimuli. Our method can discriminate components of the visual hierarchy based on their unique dynamics during various stimuli, providing an understanding of the local characteristics of neural activity directly from data, without supervision or modelling. Our approach opens new avenues to investigating non-stationary neural data and better understanding how components of the brain change in response to its environment.